Efficient Decoding of Large-Scale Neural Population Responses With Gaussian-Process Multiclass Regression.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Standard neural decoding methods struggle with overfitting and scalability in high-dimensional neural data.
• Understanding information content and correlation limits in neural population codes is crucial.

Purpose of the Study:

• Introduce a novel Gaussian Process Multiclass Decoder (GPMD) to overcome limitations of existing decoding methods.
• Develop a method suitable for decoding continuous variables from high-dimensional neural activity.
• Assess the role of neural correlations in population coding.

Main Methods:

• The GPMD is a multinomial logistic regression model with a Gaussian Process prior on decoding weights.
• Incorporates hyperparameters for automatic pruning of uninformative neurons.
• Utilizes variational inference for efficient fitting, scaling to thousands of neurons and outperforming in low-trial regimes.

Main Results:

• The GPMD achieves state-of-the-art decoding accuracy in primary visual cortex recordings across monkeys, ferrets, and mice.
• Demonstrates substantial performance improvement over independent Bayesian decoding.
• Confirms the essential role of correlation structure for optimal neural decoding.

Conclusions:

• The GPMD offers a scalable and accurate solution for neural decoding from high-dimensional population activity.
• Highlights the critical importance of considering neural correlations for effective information extraction.
• Provides a robust platform for analyzing neural population codes across species.